NPS ARCHIVE

1966
DIETRICH, W.

^881
CALCULATIONS OP THE EFFECTS OF PEAK
CLIPPING ON SPEECH-LIKE SIGNALS

WILLIAM YARNER DIETRICH

Jmttmmii

I 1 ffi u188uffl H
H 18 ffl H fi I fin ifniul
, POSTGftfltJUATJSSCHOOL
SREY, CALIF. 93940

DUDLEY KNOX LI6RARV

NAVAL POSTGRADUATE SCHOOL
MONTRREY CA 93! ^-5101
This document has been approved for public
Release and sale; its distribution is unlimited.
DUDLEY KNOX LIBRARY
NAVAL POSTGRADUATE SCHOOL
MONTFREY CA 93943-5101

CALCULATIONS OF THE EFFECTS OF

PEAK CLIPPING ON SPEECH-LIKE SIGNALS

by
William Varner Dietrich
n
Lieutenant, United States Navy
B.S., University of Wisconsin, 1962

Submitted in partial fulfillment

for the degree of

MASTER OF SCIENCE IN ENGINEERING ELECTRONICS

from the

NAVAL POSTGRADUATE SCHOOL

December 1966
 ABSTRACT
Peak clipping is a well known method of increasing the average

power output of a peak power limited voice communication transmitter.

Although the clipping process introduces distortion, articulation tests
have shown that clipped speech remains highly intelligible. Using
idealizations of vowel sounds based on the mechanism of speech pro-
duction, calculations were made of the spectra resulting from clipping

these speech-like signals. The results indicate a high degree of

similarity between the spectra before and after clipping. The power
gained by clipping at audio frequency and at narrowband was calculated

and compared with previously published data. Repeaking due to com-

ponent rejection was investigated for clipping at audio and narrowband.
Calculations of the effect of varying the phase characteristic of the

signals before clipping indicate that such variation may improve the

intelligibility of clipped speech.

 LtlHARY
NAVAL POSTGRADUATE SCHOOL
MOTfPEPEY, CfcUF. 93940
TABLE OF CONTENTS
Section Page
1 . Introduction 7

2 . The Nature of Speech 9

3. Vocal Tract Analog 11

4. Speech Spectra and Intelligibility 14

5. Intelligibility of Clipped Speech 17

6. The Clipping Process 26

7. Results of Calculations 32

8. Repeaking 49

9. Effects of Phase Characteristics on Clipping 52

10. Conclusion 55

11. Bibliography 56

Appendix
A. Spectra of Clipped Signals 59

B. Fourier Analysis 102

C. Digital Computer Programs 103

 LIST OF TABLES
Table Pa 9 e

I. Amplitudes of Spectral Components Resulting

from Clipping a Signal of the Form
A cos oJ t + B cos Jdt 28

II. Coherence Coefficients after Clipping 41

lit; Results of Power Calculations 42

IV. Comparison of Power for Gradual and

Abrupt Clipping 47

V. Comparison of Coherence Coefficients for

Gradual and Abrupt Clipping 47

VI. Effects of Original Phase Characteristic

on Clipping 53
 LIST OF ILLUSTRATIONS
Figure age
1 . Block diagram of voice mechanism 9

2. Electrical analogs of a short length of lossless

acoustic tube 11

3.3. Articulatory analog 12

4. Idealized spectrum of the vowel sound in "bed" 14

5 . Block diagram of a channel vocoder 16

6 . The differentiating and integrating circuits used

in the articulation tests £9

7. Articulation scores for each of the ten arrangements

of the distorting circuits 2G

8. Intelligibility scores for audio clipping 22

9 . Intelligibility scores with and without differentiation

prior to SSB clipping 24

10. Comparative intelligibility for pre-modulation and

post-modulation clipping 25

11. Clipper characteristic 26

12. Relative amplitudes of the spectral components

resulting from infinitely clipping signals of the
form cos 2l>-mf t + cos 2Tr (m+l)f t 29
o o
13. Relative amplitudes of the spectral components
resulting from infinitely clipping signals of the
form cos 2lVmf t + cos 2iY(m+l)f t +
°
cos 2 ir(m+2)f ? 30
o
14. Amplitude and phase characteristics for the vowel
sound in "bed" 35

15. Spectrum of "bed" clipped 10 db 36

16. Spectrum of "bed" clipped 20 db 37

17. Spectrum of "bed" infinitely clipped 38

18. Spectrum of "bed" infinitely clipped at SSB 39

19. Average in-band power increase with clipping 44

20. Total power increase with clipping 44

21. In-band power vs . coherence coefficient 45

22. Gradual clipping characteristic: y = tanh x 46

23. In-band power increase for gradual and abrupt
audio clipping 48

24. Time waveforms for "bed" at audio and SSB 51

25. Time waveforms for "bed" with all components

in phase 52
1 . Introduction

Although modern technology has provided the means to transmit

large amounts of information efficiently, conventional voice communi-

cation systems have retained wide usage. While they are comparative-
ly very inefficient, conventional speech communication systems will

probably remain attractive for many applications because of their

simplicity. An important factor contributing to the inefficiency of

voice systems is the unsuitability of the speech signal for transmission

by conventional means -- amplitude modulation or single sideband

systems .

Speech waveforms characteristically have large peaks relative to

their rms value. The peak factor, defined as

PF = 20 log peak value ,

rms value

of speech is typically about 14.5 db. [14] In a conventional peak

power limited transmitter, the amplitude of the modulated signal must
be constrained so that the transmitter is not over-driven on the signal

peaks. The average power in the speech signal being transmitted, then,
is only a small fraction of the transmitter's peak power capability. For

example, in a system where the information-bearing power is limited to

100 watts, a speech signal having a 14.5 db peak factor would have an

average power of only 3.65 watts. Work done in speech processing for

conventional transmission, therefore, has been focussed on methods of

reducing the peak factor of speech waveforms.

The method most widely used is perhaps the most straightforward

one: merely limit, or clip, the peaks of the speech signal. The clip-
ping process, to be discussed further in Section 6, produces distortion.

As the degree of clipping is increased the peak factor is reduced, but

the amount of distortion generated increases. The trade-off between
reduction of the peak factor and the increase in distortion is difficult

to optimize, because it is not known how these two effects interact to

influence the intelligibility of speech in noise.

 The succeeding sections contain a brief description of the nature
of speech and an electrical vocal tract analog. Following this is a

discussion of some of the factors affecting speech intelligibility and a

summary of the results of experiments conducted to determine the in-

telligibility of clipped speech.

In Section 7 a model of a speech signal is proposed for the purpose

of calculating the spectrum which results when this speech-like signal

is clipped. The computed spectra are presented and the reduction in

the peak factor obtained is compared with previously published results.

Repeaking of the clipped signals due to filtering and frequency trans-

lation is computed in Section 8. Finally, the effects of varying the

phase characteristics of the speech-like signals on the clipping process

are analyzed.
2 . The Nature of Speech
A speech sound begins with a flow of air under pressure from the

lungs through the trachea to the larynx. At the larynx the air stream

may be periodically interrupted by the vibration of the vocal cords to

produce a voiced sound, or it may pass through uninterrupted producing

an unvoiced sound. In the case of a voiced sound, the periodic pulses

of air produced by the vibration of the vocal cords excite the resonant
modes of the vocal tract cavities to produce a distinctive sound. As
the configuration of the vocal tract is changed, (by movement of the

throat, tongue, and jaw, etc.) its acoustic transfer function changes

so that different sounds are produced.

Unvoiced sounds are generated by forcing the air stream through

constricted openings such as between the tongue and the upper teeth.

These noise-like sounds are also modified by the resonances of the

vocal tract. The mechanism for producing speech sounds is shown in

block diagram form in Figure 1 . [6 ]

NASAL
-NOSTRILS
CAVITIES

ORAL
LUNGS TRACHEA LARYNX THROAT TEETH LIPS
CAVITIES

Fig. 1 Block diagram of voice mechanism

During normal breathing the vocal cords, which are actually two

thick folds or bands, are widely separated at one end, forming a large

triangular opening. To produce a voiced sound the cords are drawn to-
gether, closing off the air passage except for a thin slit. The act of
exhaling then sets the vocal cords vibrating; the slit opens and closes
periodically. The fundamental frequency of vibration varies from about
90 Hz for a deep voiced man to a maximum of about 350 Hz for a high-
voiced woman, with typical values of 125 Hz and 250 Hz respectively [6] .
The vibration is by no means purely sinusoidal since the slit may remain
closed for as long as half of the cycle. The sound spectrum produced
therefore contains a fundamental and many harmonics. R. L. Miller

has shown that the relative amount of energy present in the higher
harmonics is related to the abruptness of closing the vocal cord slit.

A speaker may increase the amplitudes of the higher harmonics by

causing the slit to close more abruptly. [12] Speech sounds may have
significant energy resulting from vocal cord action at frequencies as

high as 40 times the fundamental frequency. [6]

The fundamental frequency, or pitch, of a voiced sound is changed

by varying the tension of the vocal cords. As tension is increased

the cords become firmer, and they are stretched to a greater length.

The result is an increase in fundamental frequency.

Voiced sounds derive their distinguishing characteristics from the
configuration of the vocal tract, the throat and oral cavities. Corres-
ponding to a given configuration of the vocal tract there is a certain

set of acoustic resonances which selectively transmit the harmonics of

the sound source. The spectral regions of reinforcement are called

formants . The action of the vocal tract in reinforcing some frequencies
while attenuating others may be thought of as impressing a type of

modulation on the source signal, or carrier.

The nasal tract acts in the same way to modify the source spectrum,
but its configuration is essentially fixed.

10
3. Vocal Tract Analog

The mechanism for speech production, composed of an active

part -- the sound sources — and a passive part — the vocal and nasal
tracts, has led to the development of an electrical analog of the vocal
tract. The vocal tract is viewed as an acoustic tube of varying cross-
sectional area. If the cross-sectional dimensions are small compared

to a wavelength of the sound, no error will be made by considering the

tube to have a circular cross-section. The electrical analog of a
cylindrical tube is a transmission line where current is analogous to

volume velocity and voltage is analogous to sound pressure. [22]

The characteristic impedance at any point on the transmission line

analog depends on the cross-sectional area at the corresponding point

on the acoustic tube. The distributed parameter transmission line is

approximated by a series of lumped parameter sections where each

section represents a given length, 1, of the distributed line. The
approximation is valid at frequencies where 1 is small compared to a
wavelength. This condition, as well as the one concerned with cross-

sectional dimensions, is satisfied at frequencies below 5 kHz. [22]

The lumped parameter sections may be realized in the form of P ,

TT or sections as shown in Figure 2.

O
1i
Tram 1

TT ^= •

)
'Z.

o —a
4
o tt>~
*Z£-~
T

Fig. 2 Electrical Analogs of a short length

of lossless acoustic tube.

11
The relations for the electrical parameters in terms of the dimensions

of the acoustic tube are [23] :

/>/ ^ = kAi
L = C
kA
fie
where:

A = cross-sectional area of the tube

i = length of section

c = speed of sound
= density of air
f>

k = an arbitrary constant, the ratio of acoustic to electrical

impedance
NOSE
RAOIATION
IUPEDAMCE
NASAL ANALOG
L.

HASAL
COUPLING »- feLjJ TTTTTTTT

MOUTH
VOCAL-TRACT ANALOG flAOIATION
IMPEOANCE
L. I L. L, L. L. t, * L,
——/*gfi

lc„ JL icL :

^-c „ ^c ,
:

\ ^-el
5l.c"a

* * » 4 6 4 4 1 4 4
OUTPUT
*., A.i A, A, A, »4 C,L, L, C,/C,
-
COMFICUrtATlO.S COSTROL S;G. 4ALS

Fig. 3 Articulatory analog.

The lumped sections are connected in cascade to form an analog of an

idealized vocal tract made up of approximately 35 cylinders, each |cm

long, placed end to end. When excitation is added the result is a

complete articulatory analog shown in Figure 3.

Voiced sound excitation is provided in the analog by a "buzz

generator" which produces a periodic train of pulses, narrow enough

so that the amplitudes of the harmonics are nearly constant over

the audio frequency range. The buzz filter attenuates the excitation

12
so that the spectrum decreases at six db per octaize with increasing

frequency. The additional high frequency de-emphasis dee to the

output connection into the vocal tract analog produces a spectrum

envelope that decreases at 12 db per octave.

For unvoiced sounds, excitation is provided by a gaussian noise

source having a uniform spectrum extending from 100 to 10,000 Hz.

Davenport has shown experimentally that the Gaussian distribution

is a close approximation to the amplitude distribution of unvoiced

speech sounds. [3] The noise source may be inserted at various points
along the vocal tract analog according to the requirements of the sound
being represented.

Speech synthesizers based on this type of an articulatory analog

have produced good quality synthetic speech. Early models were most

successful in producing vowel sounds, but more recent synthesizers,

employing digital computers to control the vocal tract parameters, have

been able to generate almost all of the elementary speech sounds as
well as whole sentences. [23]

13
 • .

4 . Speech Spectra and Intelligibility

The mechanism of speech production has naturally led researchers
to characterize speech sounds by their frequency spectra. Voiced
sounds, especially the vowels, are characterized by the location of
their formants . Although the formant frequencies for a given vowel
may vary from one speaker to another, their relative positions are

similar. The principal formants, typically three in number, almost

always occur at frequencies below 4 kHz. Figure 4 shows an idealized

spectrum of the vowel sound in the word "bed".

f.
r-
r
CO
* •*

i
V
/

> \
\
\
2 30 *
i \
ui '

\
2 j
i

-no Hi
UJ
V
1 N

>
—
h-
20 '
IS
!
**
-M"
h"
N
\
\
S

<r.
_i
|

|
1
X
1

>
—* \
\
ijj :
!

I
i \
CL 10 -
i

1 1 1 !
1 1

«
\
|

1 I
1

Xs
,

i
1

i |
I
i 1
;

.. i
i . -*-i. 1 -j— i
, 1
!

.5 1.0 1.5 2.0 2/1 i.o

FReouewcv - kHz.

Fig. 4 Idealized spectrum of the vowel

sound in "bed"

Unvoiced sounds have much less energy than voiced sounds due
to the absence of vocal chord vibration. Their energy is spread over

a broader frequency range, often extending to 8 kHz, and a distinctive

14
formant structure is usually not present.
Researchers have concluded that the intelligibility of speech
will remain good if the general shape of the power spectrum is pre-

served. This view has been confirmed by the success of vocoders in

transmitting intelligible speech. The vocoder (for VOice CODER)

represents an attempt to reduce the information rate, or bandwidth,

necessary to transmit speech. Although there are many different types

of vocoders, they are all basically analysis-synthesis schemes.
In the channel vocoder the analyzer takes the form of a bank of
band-pass filters whose outputs, when rectified and smoothed, rept->

resent the envelope of the short-time spectrum. When the voiced-

unvoiced detector indicates a voiced sound, the pitch extractor gen-

erates a signal proportional to the fundamental voice frequency.

The synthesizer is very similar in function to the articulatory

analog discussed earlier. The filter signals are used to control the

frequency response of a time-varying filter, realized as a bank of

modulators followed by band-pass filters, so that the output spectral

envelope corresponds to the one measured by the analyzer. The filter

is excited by either a pulse generator for voiced sounds, or a white

noise generator for unvoiced sounds . The frequency of the pulse

generator is controlled by the pitch signal from the analyzer.

Other direct evidence that speech intelligibility is related to

the preservation of the power spectrum comes from a system which

measures an index of correlation between the patterns of the running

power spectra of a distorted signal and the original undistorted speech
signal. [20] The "pattern correspondence index" was found to have
a direct relationship, by means of a calibration curve, to articulation

test scores for various types of distortion, including peak clipping.

15
 VOCODER MULTIPLEXING VOCODER
ANALVZER AND SYNTHESIZER
TRANSVlSS'ON

BANOPASS RECTI- LOW- PASS MO0U- BANOPASS

FILTERS FIERS FILTERS LATORS FILTERS

771 J ?oo- !
,

300 HZ '\ j
300 HZ •

_J 300-
450 HZ
MICRO- LOUD-
PHONE SPEAKER

"ft
SPEECH REMADE
SPEECH
2800- !

3200 HZ I

-*t SWITCH

PITCH PULSE NOISE

DETECTOR J GENERATOR I |
GENERATOR I

! I

Fig. 5 Block diagram of a channel vocoder. [19]

16
5. Intelligibility of Clipped Speech

Studies of the intelligibility of clipped speech began near the end

of World War II as a result of experiments conducted to determine the

effects of overload distortion on intelligibility. It was found that

speech remained at least moderately intelligible, although the quality

suffered, no matter how much amplitude limitation was introduced.
An experiment conducted at Harvard University in 1944 under the

direction of J.C.R. Licklider indicated that the intelligibility of speech

in ambient noise remained essentially constant at a word articulation

score of 80 percent as the amount of peak clipping was varied from
to 18 db. With 20 to 22 db clipping, articulation scores dropped to
about 70 percent. [13] The subjective observations on the quality
of the clipped speech noted were: a "sharp" or "scraping" effect

due to the high frequency components generated in the clipping process

and a monotonously uniform intensity. In addition it was found that

with large amounts of clipping, noise picked up by the microphone

was quite intense during the gaps between words.

Experiments at low listening levels near the threshold of aud-

ibility indicated that with clipped speech, reception was more uniform

than with undistorted speech. Without clipping, intense words were

audible but weaker ones were missed, but with clipped speech all of

the words became audible at the same level. A closely related effect

was the more nearly uniform audibility of the component sounds of

test words under conditions of severe clipping. When undistorted

speech was barely audible, the listener heard only the vowel sounds.
However, when severely clipped speech wasiheard near the threshold
of audibility, the consonants were as audible as the vowels. [13]
The Harvard report found that the quality of speech was reduced
"surprisingly little" by peak clipping. The experimenters observed
that:

1) 6 db peak clipping is barely detectable;

17
 2) 12 db
peak clipping is not at all objectionable, but,
on the contrary, sounds as though the speaker were
enunciating with special care;

3) 18 db peak clipping makes speech sound somewhat

sharp and rasping but less unnatural than speech
over a throat microphone;

4) 24 db clipping leaves speech quite intelligible but

makes it sound unnatural and "grainy". [13]

In a following study, Licklider and I. Pollack investigated the

effects of differentiation and integration in combination with infinite

clipping on speech intelligibility without noise. [10] (Infinite

clipping implies that the speech signal is amplified by a very large

factor prior to being clipped so that the output of the clipper is a

series of positive and negative rectangular pulses of equal amplitude

The only characteristics retained from the original waveform are the
times of zero crossings.) Integration and differentiation were accom-
plished by the circuits shown in Figure 6. The circuits were driven
by low impedance sources into high impedance loads. Over the fre-
quency range of interest in speech, these circuits act as "spectrum

filters" . As can be seen from the magnitude of the voltage transfer

function, the differentiator (Figure 6a) attenuates the signal less as

frequency increases; it tilts the spectrum upward at a rate of 6 db

per octave. In similar fashion, the integrator tilts the spectrum

downward, introducing an additional 6 db attenuation for each octave

increase in frequency.

Articulation tests were run with ten different arrangements of

integrator, differentiator, and infinite clipper. The results are

plotted in Figure 7. The articulation scores fall into four groups.
The first group is made up of the three tests where no clipping was
done. Although the intelligibility scores for all three tests were
near 100 percent, the observations on quality were very different.

Differentiation, because it emphasizes the higher frequencies, made

the speech sound overly crisp. On the other hand, integration made
it sound muffled and "boomy".

18
 10 fcfl.

R
r
!wd
V,
R+
F-
R+ ^
UJ
= -J
S
I0 + 5«*>

/v U) U0 < 10 <v loo uO > IOOO

v,

a) differentiator b) integrator

Fig . 6 The differentiating and integrating circuits

used in the articulation tests

The second group of intelligibility scores comes from the two

arrangements where clipping was preceded by differentiation. The
two sets of scores are almost the same, and both are always over
90 percent. Integration after clipping, while it did not affect in-

telligibility, did improve the quality.

The third group consists of the three cases where infinite clip-
ping was the initial distortion: clipping alone, clipping plus

differentiation, and clipping plus integration. In this group again

it was found that the process following clipping had little effect on

intelligibility. While integration after clipping again improved the

quality, differentiation made the clipped speech sound even worse.

The final group of curves is the pair for which integration pre-

ceded infinite clipping. The articulation scores for this group are so

19
 100 . J__ 100
• ••
•
NO DISTORTION
90
3~* ••'•- •••
100 •
» • • -
. *
•
T -
• •

DIFFERENTIATION
90
..:«• ^jj—i^j
100
****** ; i. *

INTEGRATION
o 90
h- mo
< 00
^" •
_J
Z>
o
J-
£C
<
90

•
^S^
>^^ m
• •

90

3s^
•
•
•

• •
•
•

80 < •
80
•
••J
• •

Q •
•
nr
o fO
70
5 INFINITE CLIP. INT.
•
PEAK CLIPPING
H- 60 (•CUP.)
Z
UJ
o 50
•

cr 100 DIFF CLIP. INT.

UJ
a. T T

30 •
30
•

20 •
•

•
•
•
10

•
INT. CUR INT. CLIP. DIFF.
n _i I i

5 10 15 20 25 5 10 15 20 25

SUCCESSIVE TEST SESSIONS

Fig. 7 Articulation scores for each of the ten arrangements

of the distorting circuits [10] .
low that attenuating the higher frequencies and then clipping evidently
distroys some of the important cues for intelligibility.

Additional work, was done to try to determine why the intelligi-

bility scores showed such marked improvement during the course of

the experiment. Articulation tests were run using unfamiliar words

with distortions of clipping and clipping plus differentiation. Scores

on the new tests were about ten percentage points lower than the
corresponding scores using the familiar words. However, scores for

the new words were 15 points higher than the average of the first

five original test sessions. The researchers concluded that, although

familiarity with the test vocabulary (250 words) was an important fac-
tor, the listeners acquired a general ability to identify words correctly

in spite of severe distortion.

A more recent study conducted at Montana State College in 1962

investigated the application of peak clipping to single sideband

communications systems. [21] Intelligibility tests were run with

signals which were clipped at various stages in the single sideband

modulation process and then mixed with gaussian noise. The signal-
to-noise ratio defined for these tests was:

X = 20 1og _|E_
10
n

where E = clipped signal peak value

E = rms valud of the noise

n

Signal and noise have the same bandwidth

This definition is particularly appropriate to evaluation of processing

techniques for peak power limited systems.

Figure 8 shows the intelligibility scores resulting from audio

clipping followed by low-pass filtering for a 5 kHz bandwidth. In

this case, differentiation prior to clipping did not result in improved

intelligibility. This may not, however, contradict Licklider and

Pollack's data on clipping without noise. These results will be

21
 15 Z\

Fig. 8 Intelligibility scores for audio clipping [21].

22
discussed further in Section 7. Figure 9 shows the effect of differ-

entiation prior to clipping the single sideband signal. These results

give no clear indication that differentiation either reduces or en-

hances intelligibility. -

The most significant results from the Montana State study are
those comparing pre-modulation and post-modulation clipping for

single sideband transmission. The scores plotted in Figure 10 show

that, in every case, clipping the single sideband signal provided

improvement in intelligibility over clipping the audio signal before

modulation. For both cases, A was defined at the output of the

SSB filter.

23
 iZ. & Z* 30 36
ssb Clipping lev^l (Ai)

Fig. 9 Intelligibility with and without differentiation prior to

SSB clipping [21].

24
>

-J
111

* ssb cupping X » at 4\»

o A* \5 «ib

& & AUDIO CUPPtMG >=21<JI>

u w
G> O \*\5&
M
G n
X* 9 ^V>

IZ \fc 24 30 3b 42.

CUPPING LEVEL C<M

Fig. 10 Comparative intelligibility for pre -modulation and

post-modulation clipping [21] .

25
 .

6. The Clipping Process

The clipping process may be represented mathematically by a
power series approximation to the clipper's input-output characteris-
tic. The clipper characteristic in Figure 11 is an odd function of x,

therefore its power series approximation will have only odd-order

terms of the form:

y = ax
1
+ ax 3
v5
+ ax 5 + ...,
i

x * X
max

Fig. 11 Clipper characteristic

If the input x = A cos to

o
t where C< A< xmax :

3 5
y = a .
1
(A cos u>
o
t) + a„
3
(A cos
.
to
o
t) + ar
5
(A cos wo t) + .

a A
y = a n
1
A cos to
o
t + —— 3:
4
(3 cos wo t + cos 3 a)
o
t) +

(10 cos 60 t + 5 cos 3 to t + cos 5 to t) + . .

16 o o o

Thus the output contains a component at the frequency of the input,

CO , plus harmonic distortion components at odd multiples of ix) .

o o

26
 If the input is a two-tone signal of the form x A cos u) t +

B cos U) t , the output will contain, in addition to the fundamentals

and odd-order harmonics , intermodulation distortion components at

frequencies given by iu) + j u) , where i and j are positive

?

integers and (i + j) is odd. Table I lists the fifth order approxi-

mations to the amplitudes of the spectral components . If the signal

being clipped is a narrow-band signal, ( |t» - a) |«uj ) only

four of the 16 distortion components listed will fall at frequencies

near u) and u> . The frequencies of these four are:

J. Ci

2w, "i*2
i

2w 2 _a)
l

3u) - 2 u>
l 2

3 60 - 2 u)
2 1

All other distortion components will be separated from these by at

least the order of to rad/sec. If the approximation is extended to

seventh order, there will be a total of 30 distortion components, of

which only six will fall near u3 and o> . The remaining distortion
components will lie outside the frequency band of interest and may be
removed by filtering. This reduction of distortion helps explain why
speech is more intelligible when clipped at narrow-band (SSB) than
it is when clipped at audio frequency. Figure 12 shows the relative

amplitudes of the spectrum resulting from infinitely clipping a signal

of the form cos 2 TV mf t + cos 2tr (m + For an input con-
l)f t.
o o
sisting of two cosine waves, the spectrum of the clippedsignal is

of the form A nK cos 2Trnf t,n = l,2,... . (An even function clip-
o
ped symmetrically remains an even function.) In Figure 12 the

negative amplitudes indicate phase reversal of the components.

When m is small, the input is a wideband signal and the clipping

process has a marked effect on the amplitudes of the "signal" fre-

quencies -- mf and (m + l)f . (For the case m = 1, the two
o o

27
 ..

Table I

Amplitudes of spectral components resulting from clipping a signal of

the form A cos cO. t + B cos tO t.

FREQUENCY AMPLITUDE
(rad/sec)
3 2 5
co a A + 3/4 a A + 3/2 a AB + 5/8 a A

3 2 4
+ 15/4 a A B + 15/8 a AB + . .
5 5
3 5
CO a B + 3/4 a B + 3/2 a A B + 5/8 a B

2 2 4
+ 15/i a A B + 15/8 a A B + . . .

3 5 3 2
3u> 1/4 a A + 5/16 a A + 5/4 a A B + ..
1 5
3 5 2 3
3«) 1/4 a B + 5/16 a B + 5/4 a A B + .
2
2 4 2 3
2 u> ±u> 3/4 a A B + 5/4 a A B + 15/8 a A B +
1 2
2 4 3 2
u) ±2 u> 3/4 a AB + 5/4 a AB + 15/8 a A B +
2 3
5
5u>'1 1/16
"' a cA + . . .
5

>2
1 / 16a 5 !

4
4(J, ±<0o 5/16
' a.A B + ...
1 2 5
3
3u),
1
±2 od
*>2n
V
5/8 a„A B + ...

2u>
12 ±3 u>_ 5/8
'
a cA B
5
2 3

4
+ ...

±4a)
u>_1**"»2 5/16aa 5"
5/16 c AB + ...

28
WrVr tV^v- —,4
m = 9

i ... ..i
1 ^i^—H^iV- **tt* 1
— ~riir
r
"nrr
11
1

m = 6

-
j
'-
i^ i'
i
J V "'i '—^"
;
,
n " ''
l
— rf
-
rJ~^r —
,
II
'
1 '
'' —^^—

m = 3

JL II..,, H. Y]—'
' '
t |
— * *•
—I '
—
l ' * *i » ' ' "i I ' '
i r i
*
»

m = 1

Fig. 12 Relative amplitudes of the spectral components- resulting

from infinitely clipping signals of the form
cos 2 mf t + cos 2 (m+l)f t
o o

29
 -I . , 1. 1 t | >

,
i
,
I, .
,
,>,,.,, .r-^o^yl
FT
j
|

m=26

i, ,i
T r-rr-r npr
J. i,
, .,
i.,. r , ,..i
r -j. ,...! , i.i. i ,.i i

m=12

L ^' |ii''i"i '

'-I
1

1
11

^
m=4

HI
1,.,I.aJ.,,,
1 T ,il,, ,.ii,i r—prr —J-—r—r ^y-*-^.

m=l

Fig. 13 Relative amplitudes of the spectral components resulting

from infinitely clipping signals of the form
cos 2 mf t.+ cos 2 (m+l)f t + cos 2 (m+2)f t
o o o

30
originally equal amplitudes have been distorted so that the component

at 2f is only half as large as the one at f .) As m increases, the

o o
input gradually changes from a wideband to a narrowband signal and

a relatively simple spectral pattern emerges. The pattern repeats itself

(with reversed sign) at odd multiples of (m + £ ) f . Figure 13 shows

the same type of result for a three-tone input of the form cos 2Trmf +
o
cos 2 it (m + l)f + cos 2Tf (m + 2)f .

o o
J. M. Dukes [4] analyzed the clipping process from a statistical
point of view. He found that, for a totally random signal, infinite

clipping does not change the shape of the signal's power spectral

density function. For partially constrained signals such as voiced

speech sounds, however, the shape of the power spectral density

curve is altered by clipping. In connection with this finding, Dukes
sites the results of an investigation which showed that vowels suffer
more reduction in intelligibility due to clipping than do consonants
Sinee the short-time spectrum of a vowel consists entirely of

harmonically related components , i .

o
e . , f .

1
= if , i=l,2,3,....,
clipping will generate no new frequencies in the band of the original

signal. The distortion components will fall either at the frequencies

of the original spectral components or at higher frequencies where

the original spectrum amplitudes were insignificantly small. Thus,
distortion measurements made with simple sine waves or two-tone
signals cannot indicate the degree to which the intelligibility of a

speech signal will be degraded. Some of the distortion generated

will lie at frequencies higher than those present in the original sig-

nal, and this portion might be considered to affect intelligibility in

the same manner as corrupting noise. But much of the distortion will

simply alter the amplitudes of the original spectral components.

31
7. Results of Calculations

In order to investigate the alteration of the spectral amplitudes

due to peak clipping, calculations were made, with aid of a Control

Data 1604 computer, of the spectra of clipped speech-like signals.
The signals used were idealizations of vowel sounds described in the
frequency domain by their spectrum amplitude characteristics. The sig-
nals were considered to be periodic with a period of 1/120 sec corres-

ponding to a fundamental voicing frequency of 120 Hz. Thus the idealized

spectrum of the phomeme / / in the word "bed" could be represented by
the amplitudes of 23 spectral components covering frequencies up to

2760 Hz.
The relative phases of the components were computed by inferring
a phase characteristic from the amplitude characteristic. Considering
the vocal tract analog discussed in Section 3, a peak in the amplitude

characteristic could be considered with small error to result from a com-

plex conjugate pole-pair in the transfer function or, alternatively, from

the resonance of a single L-C section. A discussion of the errors in-

volved in this approximation is given in Reference 26.

VWW- tV9 V9*-

V,

Single section of vocal tract analog.

The voltage transfer function of the circuit shown above is;

V
2 1
V
1 s LC + sRC + 1

(1 - U) L0 + jwRC

32
The phase of V~ with respect to V is:

=
-1 u>RC
<t>
- tan
1 - UJ LC

=
-1 U>RC
- tan
2
l-u.
2
L u)
O

-1 tO RC
= - tan
60 o u>
a) to
o J

but (J RC = tt
o Qo

l
then (fr
=tan
Q
ii*y
The phase characteristic corresponding to an amplitude function having,
for example, three formants is then the superposition of three phase

functions of the type shown above, one for each resonant frequency.

The Q factor for each phase function is dependent upon the formant
frequency and bandwidth according to the relation:

BW 3db Qo

For this investigation, formant bandwidths were set at 100 Hz, a

mean value taken from data in Reference 1

In Figure 14 the amplitude characteristic for the vowel sound in

the word "bed" is plotted as a series of straight line segments con-

necting successive points representing the amplitudes of the 23 dis-

crete frequency components. The curve labeled "PHI" is the phase

33
characteristic inferred by the process previously discussed. The am-
plitude characteristic was normalized so that the corresponding time

function would have a maximum absolute value of 10. The curve labeled
"PWR" is, in effect, a power distribution curve: at any given frequency
it represents the total power (across a one ohm resistor) in the components
at that frequency and below. Figures 15, 16, and 17 show the spectra

resulting from clipping the time waveform 10 db, 20 db, and infinitely.

To get, for example, 10 db clipping the time function was amplified by

a factor of 3.16 and then clipped symmetrically at +10 and -10. The
clipping characteristic was ideally "abrupt" in that, for values of the

input between +10 and -10, the clipper output was identical to the in-

put, and for values exceeding that range the output was fixed at +10

or -10 as appropriate. In these Figures, the last point on the power

distribution curve represents the total power in the clipped wave. For

the infinitely clipped wave, the total power reaches the maximum
possible value of 100.

Figure 18 shows the spectrum resulting from infinitely clipping

not the audio waveform but the corresponding upper sideband signal.

This signal was generated according to the equation

N
W>" £,
n=l
C co.[2lKf +nf )t + 4> ]
n

where C represents the amplitude characteristic and the phase

n n
characteristic. The carrier frequency, f was chosen as 24 kHz,
,

a value high enough to give meaningful results but low enough so

that the number of computations required for fourier analysis would not

be unreasonably large. remained at 12 Hz. To save

The value of f
o
time, the clipped signal was analyzed only at the frequencies of the

23 components present in the original signal.

Similar calculations were performed using models of three other

vowel sounds, the phonemes /a/, /£/ and /u/. The amplitude char-
acteristics for these models were taken from spectrum envelopes

34
-o
i
i

<

4 4 8 JO 1^
FR.£C?'J.£MCS> - kHz
Fig. 14 Amplitude and phase characteristics for the vowel sound in "bed".
35
 s

PWR

1£DB_
21 4 6 e. to /2
FRCQUSNCV -kHz
Fig. 15 Spectrum of "bed" clipped 10 db
3fi
 9

&

PWR

a:
L'j

y-

7}

ft

a*
<

01

4 6 10 /2
FfXQUtUCV - kUz
Fig. 16 Spectrum of "bed" clipped 20 db
37
 £ 4 6 6 10
rRCQUENCV - kWz
Fig. 17 Spectrum of "bed" infinitely clipped
38
 B

pU>!

a:

3
o

»0 5.

p. 3

t/1

BED

Z A 6 8 10 it

Fig. 18 Spectrum of "bed" infinitely clipped at SSB

39
calculated by G. Fant. [5] The spectrum envelopes were mathemat-

ically synthesized in terms of elementary resonance curves, one for

each formant. The process is similar to that previously discussed in

connection with inferring a phase characteristic from the formant fre-

quencies . For his calculations, Fant also assumed a constant formant
bandwidth of 100 Hz. The spectra resulting when these sound models
were clipped are presented in Appendix A.
In addition the same calculations were made with the signals
differentiated prior to clipping. Instead of differentiating the time

waveform numerically, the differentiation was done in the frequency

plane by tilting the spectrum. The spectra calculated for the differ-

entiated signals are also presented in Appendix A.

Generally, the spectra of the clipped signals are quite similar to

the orignal spectra. In particular, the formant frequencies are not al-

tered by clipping. In all of the spectra of the clipped signals, smaller

peaks appear between the formants where the original spectra had
smooth "valleys." The spectra of the signals clipped at SSB are, as

expected, noticeably "cleaner" than those where clipping was done at

audio.

To get a quantitative measure of the degree of similarity between

the shape of the power spectrum before and after clipping, a "coherence

coefficient," X, was defined as follows:

1
N
CT.
1
(T
2
£-
i=l
li 2i

where C and C , i = 1 ,2, .... ,N, are the discrete values of the

power spectra before and after clipping, respectively, and •' •

2
07 - L£
i=l
c
J1
2
, j = i,2

For each computation the value of N was set so that only those com-
ponents at frequencies which were present in the original signal were

40
included. The higher frequency distortion components generated in the

clipping process did not enter into the calculation. In the cases where
differentiation preceded clipping, the coherence coefficient was cal-
culated using the spectrum of the differentiated signal as the original

spectrum before clipping. The coherence coefficients calculated in

the various tests are listed in Table II.

TABLE II

Coherence Coefficients After Clipping

model sound at audio at SSB

lOdb 20db inf. inf.

"bed" .9688 .8808 .8430 .8723

differentiated .9669 .9046 .8536 .8506

// .9826 .9747 .9725 .9694

differentiated .9920 .9818 .9744 .9907

/a/ .9623 .9238 .9072 .9320

differentiated .9816 .9647 .9551 .9604

/u/ .9920 .9815 .9805 .9941

differentiated .9707 .9422 .9354 .9764

Listed in Table III are the values of power calculated after various

amounts of clipping. The calculated power increases as clipping be-

comes more severe because the signal was always amplified so that

its maximum absolute value was 10. The peak factor before clipping,
averaged over the four signals, is 10.9 db. For the differentiated

signals the average peak factor is 12.3 db. In every case the differ-
entiated signal has less power than the corresponding non-differentiated

wave. This result may explain why the Montana State study found
differentiated and clipped speech to be less intelligible in noise where-

as Licklider and Pollack found that differentiation before clipping en-

hanced intelligibility without noise. While, in almost all cases, the

differentiated signals have larger coherence coefficients and therefore

41
 TABLE III

Results of Power Calculations

Audio Clipping SSB Clipping

Sound Clipping Power in band of Total Power in band -width
model level original s ignal power of original signal

"bed" 5.9 5.9

10 db 26.6 27.3
20 db 69.3 71.5
inf 92.5 100.0 75.0

differen- 3.8 3.8

tiated 10 db 17.4 18.1
20 db 55.3 59,3
inf 78.1 100.0 70.5

/€/ 7.7 7.7

10 db 46.4 47.1
20 db 78.3 83.6
inf 86.8 100.0 76.1

differen- 7.0 7.0

tiated 10 db 40.8 42.1
20 db 67.7 79.0
inf 75.5 100.0 70.4

/a/ 5.8 5.8

10 db 35.6 35.9
20 db 72.3 75.8
inf 85.9 100.0 77.4

differen- 4.8 4.8

tiated 10 db 29.7 30.3
20 db 67.8 73.5
inf 83.1 100.0 75.6

/u/ 17.3 17.3

10 db 68.1 69.0
20 db 92.0 93.2
inf 95.9 100.0 79.6

differen- 9.7 9.7

tiated 10 db 47.9 48.2
20 db 76.0 77.6
inf 91.1 100.0 72.9

42
presumably greater intelligibility, perhaps the difference in power is

the dominant factor when speech is masked by noise.

In Figure 19 the average power increase in the band of the original
signal is plotted vs. the amount of audio clipping. The higher points
represent the average power increase for the signals which were dif-

ferentiated prior to clipping; the lower points are averaged over the

non-differentiated signals. It should be noted that, for this discussion,

power increase in db is equivalent to peak factor reduction in db since

all signals have the same peak value.

The increase in power due to clipping was calculated in Reference

25 from a speech amplitude probability distribution. The result of

this calculation is plotted in Figure 20 along with the average increase

in total power of the four signals from this experiment. The other
points on the graph represent data replotted from Figure 8 by taking the

values of A for which the curves for 12 db clipping, 6 db clipping, and

no processing cross the 60% intelligibility line. For example, with no

processing 60% intelligibility corresponds to a X of about 18 db and

with 6 db clipping the corresponding A is 13.5 db. The difference be-

tween these two values of A should be a close approximation to the

reduction in peak factor due to 6 db clipping

In Figure 21, the in-band power is plotted versus the coherence

coefficient. Although the relationship appears to be nearly linear, the

slopes of the curves vary widely. In the case of the model sound /a/,
data for the differentiated signal has a larger slope than for the non-

differentiated wave, but for /u/ the reverse is true. For the sake of

clarity, points were not plotted for the signals clipped at SSB, how-
ever these dafa have no consistent relationship to the curves shown.

N. W. Huddy [8] investigated the effects on intelligibility of

using a clipper having a gradual clipping characteristic instead of the

more common abrupt clipper. It was thought that the gradual clipper

might give better results since its input-output characteristic did not

have a sharp bend. At modest clipping levels, the output waveform would

43
 I

/o DIFFERENTIATED
BEFORE CUPPING
</>

a.
ID
d
o
2
-
5
cc

3
o
a.

\0 £0
CLIPPING - <U>
# eo

Fig. 19 Average in-band power increase with clipping

U /o ©
<
it)
a: EI
'O
o
2 . . CALCUOTeO FROM
Q
(B RftPLOTTeO FfcO* C»6- 8
u
© RESULTS OF THVS EXVeRl/AEMT
3
o
a.

6 I* (8 24
CUPPtNG - <ib

Fig. 20 TotaL power increase with clipping

44
 loo

© /u/

B/e/

.60

BED
A "BED" DIFFERENTIATED

Q /Q/

^ /a/ DIFFERENTIATED

.95" .90 .85 .so

COHEREMCE COEFFICIENT

Fig. 21 In-band power vs. coherence coefficient

45
have rounded rather than sharp corners and therefore should contain
less distortion. The intelligibility tests, however, showed no
significant difference between the two types of clipper when additive
noise was present.

Calculations were made using the gradual clipping characteris-

tic shown in Figure 22: the hyperbolic tangent function.

Fig. 22 Gradual clipping characteristic: y = tanh x

The results are listed in Tables IV and V alongside the corresponding
data for abrupt clipping. The gradually clipped signals have less
power (higher peak factor) and, in most cases, slightly larger co-
herence coefficients. Where direct comparisons could be made, the
coherence coefficients for gradual clipping exceeded those for abrupt

clipping by an average of .0058. The peak factors of the gradually

clipped signals averaged 1.15 db higher at 10 db clipping and 0.77

db higher at 20 db clipping. Figure 23 shows the in-band power

increase with clipping. As could be expected, the increase in power

becomes insignificantly small as for clipping levels above 30 db.
Evidently, any advantage which the gradual clipper may have
because it generates less distortion is cancelled out by the higher

peak factor of the clipped wave.

46
 TABLE IV
Comparison of Power for Gradual and Abrupt Clipping
Audio Clipping SSB Clipping

Sound Clipping In-band Total In-band

model level power power power
Abrupt Gradual Abrupt Gradual Abrupt Gradual

"bed" 10 db 26.6 21.4 27.3 21.9

20 db 69.3 56.3 71.5 58.1 55.5
30 db 84.1 87.2
40 db 92.5 99.7 74.71
inf 92.5 100.0 75.0

/a/ 10 db 35.6 26.5 35.9 26.8

20 db 72.3 62.3 75.8 64.4 62.8
30 db 79.5 86.1
inf 85.9 100.0 77.4

/a/ 10 db 29.7 22.5 30.3 23.0

differ- 20 db 67.8 56.9 73.5 60.4 57.4
entiated 30 db 76.8 86.2
inf 83.1 100.0 75.6

TABLE V
Comparison of Coherence Coefficients for Gradual
and Abrupt Clipping
Audio Clipping SSB Clipping

Sound Clipping Abrupt Gradual Abrupt Gradual

model level

"bed" 10 db .9688 .9679

20 db .8808 .8964 .9050
30 db .8525
40 db .8430 .8726
inf .8430 .8723

/a/ 10 db .9623 .9691

20 db .9238 .9323 .9452
30 db .9067
inf .9072 .9320

/a/ 10 db .9816 .9841

differ- 20 db .9647 .9670 .9677
entiated 30 db .9584
inf .9551 .9604

47
 .

— <&

-
10

6" - GRADUAL CLIPPING

© ABRUPT

20 40
VA
i
10 30 <x>

GRADUAL CUPPIW6
© A&R.UPT

VA oo
40

A/ DIFFERENTIATED
• - — • gradual clipping
© Abrupt "

YA
30 40 oO

CUPPING - Jib

Fig. 23 In-band power increase for gradual and abrupt audio

clipping

48
8. Repeaking
It is well known that when a clipped speech signal is filtered to

reduce its bandwidth the peak factor increases. One explanation for

the increase is obviously that the out-of-band power is removed by

filtering leaving, for the same peak value, a lower rms value. Another
explanation that has been advanced is that the phase characteristic of
the filter causes the components to add in such a way that the signal

repeaks. In Section 7 repeaking due to elimination of the out-of-band

power was taken into account by computing the power remaining in the

band of the original signal.

To investigate repeaking further, the spectrum of the clipped signal

was transformed back to the time domain using only the components
in the band of the original signal. The process is equivalent to passing

the clipped signal through an ideal filter having a rectangular amplitude

characteristic and a linear phase characteristic. For the case of in-

finite SSB clipping, calculations were made using the sound models
"bed" and /e/ ' . The peak values of the time waveforms increased (from

10) to 16.85 for "bed" and 16.75 for /e/. The resulting peak factor
was 5.72 db in both cases. For infinite clipping at audio followed, by

ideal filtering, the average peak factor for the four sound models was
3.17 db
These results might seem at first glance to indicate a significant

advantage for audio clipping: a smaller resultant peak factor. Upon

closer examination, however, the large difference in peak factor dis-

appears. The narrowband signal which has the smallest possible peak
factor is a simple sinusoid of constant amplitude. If this sinusoid

were to be altered so as to reduce its peak factor, for example by making

it a rectangular wave, the altered signal would contain harmonics at

multiples of the sinusoid frequency and thus would no longer be a

narrowband signal. Therefore the smallest peak factor that an SSB

signal can have is 3 db. The peak factor .computed for the clipped SSB
signals is only 2.72 db above this minimum.

49
 It has been assumed in various studies [21 , 24] that the peak
factor of a signal changes significantly with frequency translation from

audio to SSB. Three calculations were made in order to check this

assumption. When the undipped model sounds "bed", /a/, and /e/
were translated to narrowband, the peak factors changed only +0.2 7 db,
+0.65 db, and -0.03 db respectively. This is a rather surprising result

considering that the frequency components which are harmonically re-

lated at audio are of nearly the same frequency at narrowband. When

two of the clipped and filtered SSB signals, "bed", and /&/ ', were
translated back to audio, the peak factor changed by +0.01 db and -0.48

db respectively. Shown in Figure 24 are the time waveforms for "bed"

at audio and SSB.

50
 o
CD

la?

is

PQ
CO
CO

id

Amplitude

51
9. Effects of phase characteristics on clipping

Calculations were made to determine what effect the original phase

characteristic has on the clipping process. The four model sounds were
infinitely clipped at audio with: 1) their components all in phase,

2) the phase characteristic inferred from the spectrum amplitude dis-

tribution, and 3) the phase angles determined by a random number gen-

erator so that they were uniformly distributed from to 2rr . The

results of the calculations are shown in Table VI.

The signals have the largest peak factors before clipping when
their components are all in phase. (PF = 20-10 log l0 (Power) db since

the signals were normalized so that their maximum absolute value was
10.) This result was expected since the expression

N
f(t) = 5""*
*--i
n=l
C
n
cos (2 1> nf
on
t + (b )

has an absolute maximum at t = when the (b 's are all zero. The
n
coherence coefficients for the in-phase signals are all lower than for

the same signals with other phase characteristics, probably because the
large positive peak causes much of the signal to be below the zero axis.
(See Figure 25) When the signal is clipped, information is lost because

this portion of the wave has no zero crossings.

<&} m <m -
'.-Wo rwrfseO

Fig. 25 Time waveform for "bed" with all components in phase

52
 TABLE VI

Effects of Original Phase Characteristic on Clipping

Sound Power before In-band power Coherence

model clipping after clipping coefficient

"bed" Inphase 4.7 45.4 .7461

Phase inferred 5.9 92.5 .8430
Random phase 15.2 89.4 .9244
12.4 83.9 .8698
22.1 93.7 .9042

/€/ Inphase 3.4 88.4 .9193

Phase inferred 7.7 86.7 .9725
Random phase 11.8 83.7 .9844
19.4 83.0 .9857
18.5 85.3 .9798

/a/ Inphase 3.2 84.5 .7665

Phase inferred 5.8 85.9 .9071
Random phase 18.1 87.3 .9003
21.2 86.3 .9433
19.3 88.0 .8998

/u/ Inphase 9.8 93.6 .9375

Phase inferred 17.3 95.9 .9805
Random phase 18.7 95.2 .9879
22.8 95.6 .9840
16.4 95.2 .9758

53
 With the phase characteristic inferred from the spectral amplitude
distribution, the peak factors are smaller and the coherence coefficients

larger. With random phases the peak factors decrease even further.
Since the ear is quite insensitive to phase variations [6], this result

suggests that the peak factor of speech could be reduced, without im-

paring intelligibility, by passing it through some non-linear phase

network which would not distort the spectrum appreciably. For any sig-

nal expressed in the form of Eq. 9.1, there is some phase characteris-
tic which minimizes its maximum absolute value. The questions of

what that phase characteristic is, and whether or not an approximation

to it can be obtained with a realizable network are subjects for further

research in this area. Notice also that the coherence coefficients are
generally larger for the random phase signals. This indicates that the

intelligibility of clipped speech might be enhanced by some phase

variation before clipping.

54
10. Conclusion
The analysis of the effects of peak clipping on speech-like signals
has produced three important findings. First, the coherence co-
efficients are large, indicating a high degree of similarity between the
power spectra before and after clipping. Thus it is not so surprising

that a speech signal subjected to the severe amplitude distortion of

infinite clipping can be understood. Second, the peak factor did not
change appreciably with frequency translation in two specific instances:
the translation of an undipped audio signal to narrowband and the trans-
lation of an infinitely clipped, ideally filtered signal back to audio.
Third, it may be possible to reduce the peak factor of unprocessed
speech significantly by alteration of its phase characteristic. An
appropriate variation in phase before clipping may also enhance the

intelligibility of clipped speech by increasing the coherence between

the spectra before and after clipping.

Although experimental data on actual vowel sounds by themselves was

not available for direct comparison, the calculated results are in gen-

eral agreement with published figures. The real measure of effective-

ness of any speech processing system is in terms of intelligibility of

the processed speech in noise, but the method of analysis used in this

study can give indications which point toward improved speech pro-

cessing schemes.

55
 BIBLIOGRAPHY

1 . Bogert, B.P. "On the Band Width of Vowel Formants," Journal

of the Acoustical Society of America , v. 25, (July 1953), 791-792.

2. Chang, S., Phil, G.E., and Essigmann, M.W. "Representation

of Speech and Sounds and Some of Their Statistical Properties,"
Proceedings of the Institute of Radio Engineers , v. 39, (February
1951), 147-153.
o
3. Davenport, W.B. "An Experimental Study of Speech Wave Prob-
ability Distributions, " Journal of the Acoustical Society of America ,

v. 24, (July 1952), 390-399.

4. Dukes, J.M.C. "The Effect of Severe Amplitude Limitation on

Certain Types of Random Signal; A Clue to the Intelligibility of
'Infinitely' Clipped Speech," Proceedings of the Institute of
Electrical Engineers (London) v. 102-103, Pt. C, 1955-56, 88-97.
,

5. Fant, G. Acoustic Theory of Speech Production , *S Gravenhage:

Mouton & Co. , 1960.

6. Fletcher, H. Speech and Hearing in Communication . New York:

D. Van Nostrand Company, 1953.

7. French, N., and Steinberg, J. "Factors Covering the Intelli-

gibility of Speech Sounds," Journal of the Acoustical Society of
America , v. 19, (January 1947), 90-119.

8. Huddy, N.W., Jr. "An Investigation of Methods of Improving

the Intelligibility of Audio Frequency Speech in Noise." Master's
thesis, Naval Postgraduate School, Monterey, California, 1966.

9. Licklider, J. "Amplitude Distortion and Speech Intelligibility",

Journal of the Acoustical Society of America , v. 18, 1946,
429-434.

10. Licklider, J., and Pollack, I. "Effects of Differentiation, In-

tegration, and Infinite Peak Clipping on the Intelligibility of
Speech," Journal of the Acoustical Society of America v. 20, ,

(January 1948), 42-51.

11. Middleton, D. and Van Vleck, J. "The Spectrum of Clipped

,

Noise " Proceedings of the Institute of Electrical and Electronics

,

Engineers ,v. 54, (January 1966), 2-19.

12. Miller, R.L. "Nature of the Vocal Card Wave, " Journal of the
Acoustical Society of America v. 31, (June 1959), 667-677.
,

56
13. Office of Scientific Research and Development, The Effects of
Amplitude Distortion upon the Intelligibility of Speech OSRD ,

Report No. 4217, Harvard University, 1944.

14. Pappenfuss, E. Bruene,

, W. and Schenike, E. Single Sideband
,

Principles and Circuits . New York: McGraw-Hill Book Company,

1964.

15. Pollack, I. "On the Effect of Frequency and Amplitude Distortion

on the Intelligibility of Speech in Noise, " Journal of the Acous-
tical Society of America v. 24, (September 1952), 538-540.
,

16. Pollack, I., and Pickett, J. "Intelligibility of Peak-Clipped

Speech at High Noise Levels, " Journal of the Acoustical Society
of America v. 31, (January 195 9), 14-16.
,

17. Rosen, G. "Dynamic Analog Speech Synthesizer," Journal of

the Acoustical Society of America v. 30, (March 1958), 201-209.
,

18. Rostovtsev, Y.G. "The Possibility of Using Maximum Amplitude

Limitation of Speech Signals in Systems of Communication,"
Telecommunications 1958, 643-648.
,

19. Schroeder, M. "Vocoders: Analysis and Synthesis of Speech,"

Proceedings of the Institute of Electrical and Electronics En-
gineers v. 54, (May 1966), 720-733.
,

20. Schwarzlander, H. "Intelligibility Evaluation of Voice Communi-

cations ," FJ^ctronics_, (May 29, 1959), 88-91.

21. Shyne, N.A. Speech Signal Processing and Applications to

Single Side-Band . AFCRL 62-64, Montana State College, 1963.

22. Stevens, K.N., and Fant, C.G.M. "An Electrical Analog of the
Vocal Tract, " Journal of the Acoustical Society of America v. 25, ,

(July 1953), 734-742.

23. Stevens, K.N. et al "Speech Analysis and Synthesis Final Re-

port," Massachusetts Institute of Technology AFCRL 64-300,
.

December 1963

24. U.S. Army Signal Corps. Premodulation and Postmodulation Clip-

ping in Single Sideband Transmission. AEPG-Sig. 960-86, Uni-
versity of Arizona, 1960.

25. Wathen-Dunn, W. and Lipke, P. "On the Power Gained by

,

Clipping Speech in the Audio Band, " Journal of the Acoustical

Society of America v. 30, (January 1958), 36-40.
,

57
26. Weibel, E.S. "Vowel Synthesis by means of Resonant Circuits,
Journal of the Acoustical Society of America v. 27, (September
,

1955), 858-865.

58
 APPENDIX A

SPECTRA OF CLIPPED SIGNALS

59
 PHJ
-CI

ORIGINAL SP£CT£UM
lor "BCD'

rs

<

Pklft

JBEH
4 4 e >p a
FR.6QUtK!CV - kMz

60
 9

u
SPecTfcum of Beb"
DlFF6Rev)TlATgO 4 CtlTPfft <0«U

0£

3
o

io -'-

'

PUR

jam
4 6 8 10 /2
Frcouemcv - kWz

61
9

specTfcu* or "aco"

4 6 8 10 /Z
FrcQUEHCV - kHz
62
 c
l

M
SPECTRUM OF B£D"
DiFFeRCNTiM^D 4 «UPPc& imfimitcu?

'o
*6

4 6 8 10 re

_
9

SPECTfcUIA 6F "Bed" D|Frefe €W Ti/neD

QUPP£D JMFnoiTets> AT SSR

a /O /2
F 3£G.UE UCV - kHz
64
9

&

^PeCTfcUM OF "BCD"

PU/R

4 6 e 10 /2
FRCQUENCY - kU£
 c
i

z
SPECTRUM op H 5€0"
GRADUALLY CuPP^O ZO Ab

7.

PvjR
•
/
/
3
6
i

5. i

IS

V |
1

! !

, i

<
!

-J !

(/1

V
I

Hi
\

z
^ V^"^—v—
4
^-V.

FrcQUEHCV
-f,
_+.

6
- kHz
8/0
1

/2
DflTlfc

66
o

SPECTfcMNV OF "BCD"

10 /2
- kHz
 PUItf

SPecTfeu* op "Bea M
G>£M>UHJL*? CLlPP£0 so 4L
g

to

i
1 i

<
-J

2
£

!
fi

k I
I

A
A /\

AVry~A- ^<^*VW>^ .» 39Q E) -

£ 4 6 8 /O /2
rSCQUEHCV - kHz

68
G»J2-fcDUALXv> CClPPCD 40 M>

fl&DB

4 6 8 /o /E
Ffxquemcv - kHz
KQ
 B PUIS

r~

%
^PeCTfcUM OP 'g€D"

'o

a;

<

I-
! I

BEO

8 /© /Z
Frcquemcv - kU<

70
 9

I
PHI

ORIGINAL SPGCTR.UM
x £

?-

'3

c:
I-

PUIfi

., AH
4 6 a /o /2
FfcCQUfcHCV - kHz
71
 ^T>€CmUM *P /a./
p

P!-.'R

- IBDEl
10 re
o
t

PUR

^PecrKutA op /a/
CUPPE5 2oA\>

4 6 /o ft
Fpxqugmcv kU«

1A.
 VJK

SPECTRUM OF /*/

4 6 8 10 it
Frcquemcv - kHz
9

PU/R

SPecx fcuivv of /«./

iKiPiwireLV ctiP-Pen at SS>8

4
Frcqupucv
6a - kWz
/c /2

_
 OftJGlUAl SPCCTfcUM
o
for /a/
or

3
o
PL

3
U

I
a 3
<

CL
V-

C/7

PhJR

iH
4 6 8 10 it
Ff&qushcv - kU;

76
'o

4 6 8 /<? /a
pRSQUQMCV -kHz
77
 ?m

SPECTRUM OP /* (
k
DlFFCRENTlATCD { CtiPPe* Zo\\>

4 6 8 10 /z
Frcquencv - kHz
78
 r^tf-v

S?eCTfcUM OF /a./

4 6 8 10 /2
Fsxquemcy - kHz
79
s

Pu«
StfcCTfcUM OP
/«./ DlFFCfcCfoTlATFa
(MrifcHTCLtJ CUPP60 «T SSB

8 /0 /2
rPXQUENSV - k^
80
 p.

o
;6

3
o

<r

V)

<

c>

01
B

PUR

SperreuM of (*> I

-
8 10
9

PwW

^P€CTfcU/w OP /a./

£CADUALcV C(-if>p£D 3oJk»

4 6 8
Vv A
/o
/A

It
Frcqu^hcv - kHz
Q
I

1
10 /2
- kUz
84
 2<

<

w hi

3 .

a
I-

-J

<

^. t.

4 4> 8 >Q

k.H<
 *'

'SPECTftU* or /€/

g
* (>

at
Ui
O

10
5
J-

<

C-i

i*

\ \

\v V
i
v ..
v
•aJw-«-
6 8 /0
Ffjcpuemcv - kHz

86
 °>

*Spee.TfeUM. op /fc /
i

o clipped 2oJLb
* 6

S :

5
<r

4.

-J

I-

(/J

I)
" v
v -
V v 3\
2-l^JL
• -t

\ x*'-JlJO
4~ 6
?.
8 /0 tz
Ffxquemcv - kHz
87
 K-,

B\

'o
*
JMFIJOITCL^ CLlPPCfc
£;

3
o

<r

<

'

a-
C/l

Hi VV
4*8
i

u.
\\\ \\ \ Xv
/o /2
Frcqusmcv -kHz
8ft
 $

fUA

fMFlNlTeup CLIPP66 AT SS£

3
I

<
•c
a
S;

4 6s
Frcquencv -kHz
/0 It

89
s

SPCC.T*UM or /«/

4 6 8 /d /*
FrcquencV - kHz
an
 Pwk

V 1

4 6 8 /0 tt
FrcquencV - kt^
91
$

Ptofc

4 6 8 /o /z
Frcquencv -kHz
^a2
J!

ORKilKiAu SPECTRUM
W /u/

UJ
PUR

< I

nn
4 4» 8 to IX.

93
 s

PWR
J

'o

<

I-

Vli

T 8
8 10 It
Fkcquencv - kUz
94
 PUR

OP /u/
CUPPED
p

a*

3
o

,,2flm
4 6 8 10 /z
FRCQUSMdY - kHz
95
 J

'SpecTfeUM op /a /

'o

u
9
o

U)

<r
J
3

4 6 8 10 tz
Frcou&ncv - WWz
96
 a
_PU)fc

SP€crfcu* of /u/
'o /WFIWITCLV CLiW^D AT
* 6
SSE

3
o

<

J
a 3
<
-J

c-

ML
4 6 8 /0 /£

97
 t
I

Vi)
1/1

0-

OfciciMAc ^peeTfcuM

ViJ

I-

-J
6-

PUIR

aUCL
4 4 8 10 II.

FfcEQueMCV - fc.H*

98
s

^PecYfcUM OF /u/t>iFF.

PUR

4 6 8
um
10 /z
Fp.counNCV - kHz
99
4 6
FRequswcV - kHz
100
^ECTfcUM OF /tt f DJFF.

8 10 tz
 APPENDIX B

Fourier Analysis

A periodic function, f(t), with period T may be re-presented as

f(t) = -7T- +
l
n=l
/Z
—
'
f
(A
n
cos 2trnf
on
t + B sin 2-rrnf
o
t)

T o cos
2 *-, n n
n=l

where f = 1/T
o

C = A
o o
'

2 2
C
n
= J V
A
n
H B
n

_1 -B
=tan n
n
A
n .

The coefficients are determined by the following relations

T
r
A = ~| f(t) cos 2Trnf t dt

B = -r— J f(t) sin 2irnf t dt

n T J o
q

The average power in f (t) is

T °°
f A
1/T
J o
f (t)
2
dt =
-f-
+ 1/2 Z
n=l
(A
n
2
+ B
n
2
)

1/2 £
n=l
C
n

1Q2
 APPENDIX C
Digital Computer Programs

The calculations for this study were made by a Control Data 1604
digital computer with a Fortran 63 compiler. The major portion of the
computations involved transformations back and forth between the
time domain and the frequency domain. The transformations were
accomplished using the principles of Fourier analysis. The integrals
involved were evaluated by making the approximation

T N
f(x)dx - ? f (i A x) Ax where Ax = —
T
1O T~l
=1
N
1

Values of N were chosen so that the errors made by this approximation

did not significantly affect the results. Listed on the following pages

are the subroutines used in making the calculations discussed in this

study.

103
 )

SUBROUTINE ANALYZE A»Bt CtFREO.LIM)

(

C PERFORMS FOURIER ANALYSIS OF A WAVEFORM IN ARRAY S

C CONSIDERED AS A FUNCTION OF TIME (ARRAY T). A AND
C B ARE OUTPUT ARRAYS OF FOURIER COEFFICIENTS AND
C C = A**2 + B**2. FREQ IS THE FUNDAMENTAL
C FREQUENCY, LIM IS THE NUMBER OF POINTS IN ARRAYS
C S AND T.
DIMENSION A (100) ,B( 100 »C 100 »S(900 »T 900)
) ( ) ) (

COMMON/B1/S/B2/T
ARG2=2.*3. 141 59265 36*FREQ
FORM=2,/FLOATF(LIM)
29 DO 30 N=1.100
AN*0.
BN*=0.
ARG1*ARG2*FL0ATF(N
DO 35 I«ltLIM
ARG=ARG1*T( I)
A N »A N +S(I)*COSF(ARG)
35 B N =B N +S(I)*SINF(ARG)
A(N)«A N FORM
B(N)=B N •FORM
30 C(N) = ( A(N)**2+B(N)**2)
RETURN
END

104
 SUBROUTINE AN (A.BtC)
C PERFORMS FOURIER ANALYSIS OF AN INFINITELY CLIPPED
C WAVEFORM IN ARRAY S BY SUPERPOSITION OF THE
C SPECTRA OF THE COMPONENT PULSES. A AND B ARE THE
C OUTPUT ARRAYS OF FOURIER COEFFICIENTS AND
C C = A**2 + B**2.
DIMENSION A(IOO) tB(100) tAHlOO) »BI (100) ,Z(100)»C( 100)
CALL ZEROS (900,Z)
DO 10 1=1,100
10 A( I )*B( I)=0.
DO 22 1=1.100,2
IF (Z(D) 21,23,21
21 TQ=Z(I+1)-Z( I)
Tl=Z(I)+T0/2.
CALL SPECTRUM -20 . , 1 ./120. ,T0,T1 ,AI ,BI
(

DO 22 N=l»100
A(N)=A(N)+AI (N)
22 B(N)=B(N)+BI (N)
23 DO 24 I«l,100
24 C( I )=A(I)**2+B( I)**2
RETURN
END

SUBROUTINE SPECTRUM AMP ,T ,TO,U »A,B)

(

C COMPUTES THE SPECTRUM OF A PULSE OF HEIGHT AMP

C AND WIDTH X T0' CENTERED AT TIME Tl. T IS THE PERIOD.
DIMENSION A(100),B(100)
CO=AMP*TO/T*2.
*RG1*3.1415926536*T0/T
ARG2=3. 1415926536*2. *T1/T
DO 20 N=l,100
ARG=ARG1*FL0ATF(N)
C=CO*SINF(ARG)/ARG
PHI=-ARG2*FLOATF(N)
A(N)=C*COSF(PHI )

20 B(N)=-C*SINF(PHI)
RETURN
END

LOS
 SUBROUTINE ZEROS <NUM»Z)
C STORES THE TIMES OBTAINED FROM ARRAY T OF ZERO
C CROSSING OF THE WAVEFORM IN ARRAY S (NUMBER OF
C POINTS « NUM) INTO THE ARRAY Z. AFTER THE LAST
C ZERO CROSSING HAS BEEN ENTEREDt SUCCEEDING VALUES
C OF Z ARE EQUAL TO ZERO.
DIMENSION S(900)»T(900)»Z(100)
COMMON/B1/S/B2/T
J=l
LAST»1
DO 10 I«ltl00
10 ZU)«0.
DO 25 I-1»NUM
IF (S(I)) 21,22,22
21 NOW*-l
GO TO 23
22 NOW-1
23 IF(NOW+LAST) 25,24t25
24 Z(J)=T( I)
J=J + 1
25 LAST«NOW
RETURN
END

106
 SUBROUTINE SYN A.B.FREQ. TYPE »NUM)
( I

C PERFORMS FOURIER SYNTHESIS. IF ITYPE IS ZERO.

C A AND B ARE INTERPRETED AS THE CONVENTIONAL FOURIER
C COEFFICIENTS. IF ITYPE IS NON-ZERO. A IS
C INTERPRETED AS THE C COEFFICIENT AND B AS THE
C PHASE ANGLE IN RADIANS. NUM IS THE NUMBER OF POINTS
C TO BE USED OF THE A AND B ARRAYS. FREQ IS THE
C FUNDAMENTAL FREQUENCY.
DIMENSION S(900)»T(900) ,A(50),B(50)
COMMON/B1/S/B2/T
DO 21 1*1.900
S( I )=0.
ARG1=2.*3.1415926536*FREQ*T( I )

DO 20 N-l.NUM
ARG=ARG1*FL0ATF(N)
IF (ITYPE) 15.10.15
10 S( I)=S( I)+A(N)*COSF(ARG)+B(N)*SlNF(ARG)
GO TO 20
15 S( I )=S( I)+A(N)*COSF(ARG+B(N) )

20 CONTINUE
21 CONTINUE
RETURN
END

SUBROUTINE NORM (NUM)

C NORMALIZES THE ARRAY S (CONTAINING NUM POINTS) SO
C THAT THE MAXIMUM POSITIVE VALUE IS +10.
DIMENSION S(900)
COMMON/B1/S
SMAX-0.
DO 21 I=1.NUM
IF(SMAX-S( I )22. 21.21
)

22 SMAX-S( I)
21 CONTINUE
GAIN«10./SMAX
DO 23 I»1.NUM
23 S( I )=S( I)*GAIN
RETURN
END

107
 )

SUBROUTINE CLIP (CLPLVL)

c AMPLIFIES THE ARRAY S BY THE FACTOR CLPLVL AND
c CLIPS SYMMETRICALLY AT A VALUE OF 10.
DIMENSION S(900)
COMMON/B1/S
28 DO 27 1=1.900
S( I)«S< I)*CLPLVL
IF (S(I)-IO.) 24,26t26
24 IF (S(I)+10.) 25t27.27
25 S(I)"-10.
GO TO 27
26 S(I)»10.
27 CONTINUE
RETURN
END

SUBROUTINE GRADCLIP S.NUM.CLPLVL)

(

C CLIPS THE WAVEFORM IN ARRAY S ACCORDING TO THE

C CHARACTERISTIC Y * 10 * TANH( X / 10 >.
DIMENSION S(10000)
DO 10 1=1. NUM
10 S( I)*10.*TANHF(S( I )*CLPLVL/10.
RETURN
END

SUBROUTINE POWER (PWR.LIM)

COMPUTES THE AVERAGE POWER IN THE ARRAY S.
DIMENSION S(900)
COMMON /Bl/S
PWR"0.
DO 30 1=1. LIM
30 PWR*PWR+S(I )*»2
PWR=PWR/FLOATF(LIM)
RETURN
END

108
 SUBROUTINE POWERF <C,PF)
C COMPUTES THE CUMULATIVE DISTRIBUTION FUNCTION
C OF THE POWER IN THE ARRAY C.
DIMENSION C(100) »PF( 100)
PF(l)=C(l)/2,
DO 10 I«2»100
10 PF(I)»PF( I-l>+C( I 1/2.
RETURN
END

SUBROUTINE COHERE (CI »C2,NUM»COEF)

C COMPUTES THE COHERENCE COEFFICIENT BETWEEN NUM
C POINTS OF THE ARRAYS CI AND C2.
DIMENSION C1(100),C2(100)
SUM=0.
SIGSQ1«0«
SIGSQ2=0.
DO 10 I=ltNUM
SIGSQ1=SIGSQ1+C1( I)**2
SIGSQ2=SIGSQ2+C2( I )**2
10 SUM=SUM+(C1(I)*C2( I) )

COEF=SUM/ S0RTF(SIGSQ1*SIGSQ2)
RETURN
END

109
 SUBROUTINE SMOOTH (NUM»SINP #LIM)
C INTERPOLATES A SMOOTH CURVE BETWEEN NUM INPUT
C POINTS IN ARRAY C BY MEANS OF SIN(X)/X FUNCTION.
C OUTPUT IS THE ARRAY S AND WILL HAVE LIM POINTS.
DIMENSION S(9O0)fT(900)»SINP(100)
COMMON/B1/S/B2/T
INT«900/(NUM-1)
LIM-INT*(NUM-1)
SAVG=0.
DO 10 I=1»NUM
10 SAVG=SAVG+SINP( I)
SAVG=SAVG/FLOATF(NUM)
DO 11 I»ltNUM
11 SINPU )=SINP( D-SAVG
DO 20 1=1,900
T( I )=FLOATF( I)/FLOATF(LIM)/120.
20 S( I)=0.
DO 30 J=1.NUM
DO 30 I»1,LIM
ARG=3.1415926536*FLOATF I - J-l *I NT /FLOATF INT)+1.0E-50
( ( ) ) (

30 S( I )=S( I)+SINP( J)*SINF(ARG)/ARG

RETURN
END

110
 SUBROUTINE PHASE (NUM. C. PHI)
C INFERS A PHASE CHARACTERISTIC BASED ON THE
C AMPLITUDE FUNCTION IN ARRAY C. PHI IS THE OUTPUT
C PHASE ARRAY IN RADIANS.
DIMENSION S 900 ).T( 900). PHI 1(600) .C(50) ,PHI( 50),MAX( 50) .X(300).
(

* YI300)
COMMON/B1/S/B2/T
ILAST»NUM*12
DO 70 I»1.NUM
X( >=120.*FLOATF( I

70 Y(I)«C( I)
DO 71 I=12»ILAST
XINT=10.*FLOATF( I)
CALL SPLINE X. Y»NUM ,X INT.
( INT )

71 Sm-YINT
IPOS«l
J=l
DO 50 I=12,ILAST
IF (S( I+l)-S( I )46.49.49
46 IF (IPOS) 50,50*47
47 MAX(J)=I
J»J+1
IPOS*0
GO TO 50
49 IPOS«l
50 CONTINUE
JLAST=J-1
DO 55 1=1.600
55 PHIK I)=0.
DO 60 J=1.JLAST
FMAXJ=MAX( J)
Q-FMAXJ/10,
DO 60 1=1.600
FI = I

60 PHIK )=ATANF(1./(Q*(FMAXJ /Fl-FI/FMAXJ

I ) + 1.0E-20)) +PHIKI)
DO 65 1=1.50
65 PHI(I)=PHI1(12*I)
RETURN
END
SUBROUTINE SPLINE X.Y ,M. XI NT .Y INT
( )

DIMENSION X(300).Y(300) .C(4,300)

IF(X(1)+Y(M)-ATER) 10.3.10
10 CALL SPLICON(X.Y.M.C)
ATER=X(1)+Y(M)
K=l
3 IF(XINT-X(1)) 70.1.2
70 K = l

111
 ) )

GO TO 7
1 YINT*Y(1)
RETURN
2 IF(XINT-X(K+1) )6»4,5
4 YINT»Y(K+1)
RETURN
5 K«K+1
IF(M-K) 71,71»3
71 K«M-1
GO TO 7
6 IF(XINT-X(K) )13»12»11
12 YINT»Y<K)
RETURN
13 K=K-1
GO TO 6
11 YINTMX(K+1)-XINT)*<C(1»K)*(X(K:+1)-XINT)**2+C(3,K))
YINT=YINT+(XINT-X(K) * C( 2 ,K )* XINT-X K )**2+C(4 t K)
) ( ( ( )

RETURN
7 PRINT 101»XINT
101 FORMAT(8H0XINT = E18.9»32H, OUT OF RANGE FOR INTERPOLATION)
GO TO 11
END
SUBROUTINE SPLICON (X ,Y »MtC )

DIMENSION X{300>,Y(300)»C(4,300)»D(300) ,P(300) »E(300),A(300.3),B(3

100) »Z(300)
MM=M-1
DO 2 K*1»MM
D(K)*X(K+1)-X(K)
P(K)«D(K)/6.
2 E(K)"(Y(K+1)-Y(K) )/D(K)
DO 3 <«2»MM
3 B(K)«E(K)-E(K-1)
A(l,2)=-1.-D(l)/D(2)
A(1»3)»D(1)/D(2)
A(2»3)=P(2)-P(1)*A(1»3)
A(2t2)=2.*(P(l)+P(2) )-P(l)*A(l,2)
A(2.3)=A(2t3)/A(2i2)
B(2)=B(2)/A(2t2)
DO 4 K=3tMM
A(Kt2)«2.*(P(K-l)+P(K) )-P(K-l)*A(K-1.3)
b(K)=B(K)-P(k:-1)*b(k-i )

A(K,3)»P(K)/A(K t 2)
4 B(K)"B(K)/A(Kt2)
Q=D(M-2)/D(M-l)
A<M.l)=l.+Q+A(M-2,3)
A(Mt2)=-0-A(M > l)*A(M-1.3)
B M =B M-2 -A M » 1 *B M-l
(
)
(
)
(
) (

Z(M)=B(M)/A(M»2)
MN=M-2
DO 6 I=1»MN

112
Z(K)«B(K)-A(Kt3)*Z(K+l )

Z(1)«-A(1»2)*Z(2)-A( 1»3)*Z(3)
DO 7 K«1»MM
Q*1./(6.*D(K) )

C(1»K)*Z(K)*Q
C(2,K)«Z(K+1)«Q
C(3»K)=Y(K)/D(K)-Z(K)*P(K)
C ( 4 »K = Y K+l
)
( ) /D K) -Z K + l *P K)
( ( )

END

113
 I I ) ) ) ) )

C THE FOLLOWING PROGRAM ILLUSTRATES THE USAGE OF

C SOME OF THE SUBROUTINES. IT ALSO CONTAINS THE
C STATEMENTS NECESSARY TO GENERATE A SSB SIGNAL AND
C ANALYZE IT,
PROGRAM CLIPSPEC
DIMENSION S(900).T(900) ,SN(10nnO) , A 100 iB 100 .C 100 ( ) ( ) ( ) , I T ( 12 )

ORD(900) »ABS(900> »CA 100 .PHI 50 ,C1 100 »l_AB(10) ( ) ( ) ( )

COMMON/B1/S/B2/T
LAB(1)»4H10DB $ LAB 2 =4H20DB $ LAB(3)=4H INF ( )

DO 4 I-1.50
4 PHI ( I)»0.
DO 5 1=1.100
5 A( I )=B( I)=C( I )=C1( I )=0.
READ 9. LA. NUM
9 FORMAT (A4.I3)
"

READ 12. (C( I ). 1=1, NUM)

12 FORMAT (20F4.1)
DO 10 1 = 1. NUM
10 C( I )=EXPF(C( ). 230258509/2. I

C INFER PHASE CHARACTERISTIC

CALL PHASE (NUM. C. PHI)
DO 6 1=1.900
6 T( )=FLOATF( I )/900./120.
I

CALL SYN (C.PHI ,120. .1

SFIRST=S( 1)
CALL NORM (900)
GAIN=S(1)/SFIRST
DO 7 1 = 1. NUM
7 C( I )=C( )»GAIN
DO 8 1 = 1. NUM
8 Cl( I )=C(I )**2
CALL POWERF(Cl.A)
CALL POWER (PWR.900)
PRINT lOO.PWR.A(lOO)
100 F0RMATQH1.9X14H0RIGINAL POWER , 2F20 . 3/// )

DO 13 1=1.100
ORD(I )=C( I
13 ABS( )=120.*FLOATF( I
CALL ITITLE ( IT)
IT(D*8HORIGINAL
IT(2)=8H AUDIO
IT(3)=8HSPECTRUM
CALL DRAW(100»ABS.ORD»1.0.LA.IT.2000..1..0,0.2.2,7,10»0.L)
DO 135 1=1.50
135 ORD(I)«PHI( I )+8.
CALL DRAW 50 . ABS.ORD.2 .0 .4H PHI .2000.
( . 1 • .0.0.2 .2 .7. 10 .0 »L
DO 14 1=1.100
14 ORD( I )=A( I )/10.
CALL DRAWU00.ABS.ORD.3.0.4H PWR. I T.2000. . 1. .0.0. 2 .2.7. 10 .0 ,L )

IT(1)=8HCLIPPED
 CLPLVL=3. 16
DO 20 J=l*3
16 CALL GRADCLIP S , 900 CLPL VL (

CALL ANALYZF A R , CA , 1 20 . 900 ( ,

CALL POWERF (CA,A)

CALL POWER (PWR,900)
A{ 100) = PWR
CALL COHERE C 1 ,CA ,NUM ,COEF
(

PRINT 101 ,A(NUM) , PWR, COEF

101 FORMAK 10X. 17HP0WER IN SIGNAL , F2 . 3/ 1 OX 1 1H TOTAL POWER,F20.3/
*10X21HCOHERENCE CO EFFICIENT, F?0. 10///)
DO 17 1=1,100
17 ORD( )=SQRTF(CA(
I I ) )

CALL DRAW( 1 00 ABS , ORD , 1 , , LAB J ,IT,2 00 0.,l.,0,0,2,2,7,]O,n,L) ( )

DO 18 1=1,100
18 ORD( I )=A( I ) /10.
CALL DRAW( 100 ABS , ORD , 3 , , 4H PWR, I T »2000 • , 1. ,0 .6 » 2 , 2 , 7 , 10 »0 , )

20 CONTINUE
TW0PI=6. 2831853072
ARG2=TWOPI/1200000.
DO 30 1=1,10000
SN( I )=0.
ARG1=ARG2*FL0ATF( I )

DO 30 N=1,NUM
ARG=ARG1* (24000. +1 20. *FLOATF(N) )

3 SN( )=SN( )+C(N)*COSF( ARG+PHI (N

I I )

CALL GRADCLIP SN 10000 1 . ( , ,

DO 51 N=1,NUM
A(N)=0.
B(N)=0.
ARG1=TW0PI*( 24O0O.+120.*FLOATr(N) )/ 1200 000.
DO 50 1=1,10000
ARG = ARG1*FL0ATF ( I )

A(N)=A(N)+SN( )*COSF(ARG) I

50 B(N)=B(N)+SN( )*SINF( ARG) I

A(N)=A(N) /5000.
B(N)=B(N)/5000.
CA(N) = (A(N)**2+3(N)**2)
51 ORD(N)=SQRTF(CA(N) )

IT(2)=8H SSB
CALL DRAW(NUM,ABS,0RD,1 ,0,LA,TT,20 00.,l.,0,0,2,2,7,10,n,L)
CALL POWERF (CA,A)
CALL COHERE CI ,CA ,NUM ,COEF
(

PRINT 102, A(NUM),COEF

102 FORMAT( 10X9HSSB POWER , F20 . 3 / 10X2 1HC0HERENCE COEFF C ENT F20 . 10/ I I , / )

DO 52 1=1,100
52 ORD( )=A( I/10. I )

CALL DRAW( 1 00 , ABS ORD , 3 , , AH PWR , T , 2000 . , 1 . , , , 2 , 2 , 7 , 1 , , L

I

55 END

115
 INITIAL DISTRIBUTION LIST „
No. „
.

Copies
1. Defense Documentation Center 20
Cameron Station
Alexandria, Virginia 22314

2 . Library 2
Naval Postgraduate School
Monterey, California 93940

3. Commander, Naval Ship Systems Command 1

Department of the Navy

Washington, D. C. 203 60

4. Commander, Naval Electronic Systems Command 1

Department of the Navy

Washington, D. C. 20360

5. Dr. Gerald D. Ewing 3

Department of Electrical Engineering
Naval Postgraduate School
Monterey, California 93940

6. LT William V. Dietrich, USN 1

c/o Naval Destroyer School

Naval Base
Newport, Rhode Island

116
 UNCLASSIFIED
Security Classification

DOCUMENT CONTROL DATA • R&D

(Security claaalllcatlon ol tltla, body ol abatract and Indaxlng annotation muat ba antarad whan tha ovarall raport la claaalllad)

1 ORIGINATING ACTIVITY (Cotporata author) 2a. REPORT SECURITY C L ASSI Fl C A TIO*

Naval Postgraduate School UNCLASSIFIED

Monterey, California 93940 2b GROUP

3 REPORT TITLE

Calculations of the Effects of Peak Clipping on Speech-like Signals

4 DESCRIPTIVE NOTES (Typa ol raport mnd Inclualva dataa)

Master's Thesis - December 1966
S AUTHORfSj (Laat name. (Inl mini, Initial)

Dietrich, William V. LT United States Navy

6 REPORT DATE 7a. TOTAL NO. OF PACES 7b. NO. OF REM
December 1966 128 26
8a. CONTRACT OR GRANT NO. 9a. ORIGINATOR'S REPORT NUMBERfSJ

b. PROJECT NO.

9b. OTHER REPORT NOfS) (Any othar numbara that may ba aaalgnad
thia raport)

<HM
10. A VA IL ABILITY/LIMITATION NOTICES This document has been a^ roved i

release and sale; its distribution is uniimit<

_£_Al
i

+ammmam**mm*m-
11. SUPPLEMENTARY NOTES 12 SPONSORING MILITARY ACTIVITY

13 ABSTRACT
Peak clipping is a well known method of increasing the average power output
of a peak power limited voice communication transmitter. Although the clipping
process introduces distortion, articulation tests have shown that clipped speech
remains highly intelligible. Using idealizations of vowel sounds based on the
mechanism of speech production, calculations were made of the spectra re-
sulting from clipping these speech-like signals. The results indicate a high
degree of similarity between the spectra before and after clipping. The power
gained by clipping at audio frequency and at narrowband was calculated and
compared with previously published data. Repeaking due to component rejection
was investigated for clipping at audio and narrowband. Calculations of the
effect of varying the phase characteristic of the signals before clipping indicate
that such variation may improve the intelligibility of clippdd speech.

FORM
DD 1 JAN 94 1473 UNCLASSIFIED
117 Security Classification
 "

UNCLASSIFIED
Security Classification
14- LINK A LINK LINK C
KEY WORDS ROLE WT HOLE WT ROLE WT

Speech
Speech models
Clipping
Repeaking
Speech spectra
Distortion

INSTRUCTIONS
1. ORIGINATING ACTIVITY: Enter the name and address imposed by security classification, using standard statements
of the contractor, subcontractor, grantee, Department of De- such as:
fense activity or other organization (corporate author) iaaulng "Qualified requesters may obtain copies of this
(1)
the report. report from DDG"
2a. REPORT SECURTY CLASSIFICATION: Enter the over-
(2) "Foreign announcement and dissemination of this
all security classification of the report* Indicate whether
report by DDC is not authorized.
"Restricted Data" is included. Marking is to be in accord-
ance with appropriate security regulation*. (3) "U. S. Government agencies may obtain copies of
this report directly from DDC. Other qualified DDC
2b. GROUP: Automatic downgrading is specified in DoD Di- users shall request through
rective 5200. 10 and Armed Forces Industrial Manual. Enter
the group number. Also, when applicable, show that optional
markings have been used for Group 3 and Group 4 as author- "U. S. military agencies may obtain cop ies
ized,
(4) p f this
f report directly from DDC Other quaUfte##lsfs
3. REPORT TITLE: Enter the complete report title in all shall request through
capital letters. Titles in all cases should be unclassified.
If a meaningful title cannot be selected without classifica-
tion, show title classification in all capitals in parenthesis (5) "All distribution of this report is controlled. Qual-
immediately following the title. ified DDC users shall request through

4. DESCRIPTIVE NOTES: If appropriate, enter the type of

report, e.g., interim, progress, summary, annual, or final. If the report has been furnished to the Office of Technical
Give the inclusive dates when a specific reporting period is Services, Department of Commerce, for sale to the public, indi-
covered. cate this fact and enter the price, if known.
5. AUTHOR(S): Enter the name(s) of authors) as shown on 11. SUPPLEMENTARY NOTES: Use for additional explana-
or in the report. Enter last name, first name, middle initial. tory notes.
If xilitary, show rank and branch of service. The name of
the principal author is an absolute minimum requirement. 12. SPONSORING MILITARY ACTIVITY: Enter the name of
the departmental project office or laboratory sponsoring (pay-
6. REPORT DATE: Enter the date of the report as day, ing for) the research and development. Include address.
month, year, or month, year. If more than one date appears
on the report, use date of publication. 13. ABSTRACT: Enter an abstract giving a brief and factual
summary of the document indicative of the report, even though
7a. TOTAL NUMBER OF PAGES: The total page count it may also appear elsewhere in the body of the technical re-
should follow normal pagination procedures, Le. , enter the port. If additional space is required, a continuation sheet shall
number of pages containing information. be attached.
76. NUMBER OF REFERENCES: Enter the total number of It is highly desirable that the abstract of classified reports
references cited in the report. be unclassified. Each paragraph of the abstract shall end with
8a. CONTRACT OR GRANT NUMBER: If appropriate, enter an indication of the military security classification of the in-
the applicable number of the contract or grant under which formation in the paragraph, represented as (TS), (S), (C), or (U).
the report was written. There is no limitation on the length of the abstract. How-
8b, 8c, 8d.& PROJECT
NUMBER: Enter the appropriate ever, the suggested length is from 150 to 225 words.
military department identification, such as project number,
14. KEY WORDS: Key words are technically meaningful terms
subproject number, system numbers, task number, etc.
or short phrases that characterize a report and may be used aa
9a. ORIGINATOR'S REPORT
NUMBER(S): Enter the offi- index entries for cataloging the report. Key words must be
cial report number by which the document will be Identified selected so that no security classification is required. Identi-
and controlled by the originating activity. This number must fiers, such as equipment model designation, trade name, military
be unique to this report. project code name, geographic location, may be used aa key
words but will be followed by an indication of technical con-
9b. OTHER REPORT NUMBER(S): If the report has been text. The assignment of links, rales, and weights is optional.
assigned any other report numbers (either by the originator
or by the aponaor), also enter this number(s).
10. AVAILABILITY/LIMITATION NOTICES: Enter any lim-
itations on further dissemination of the report, other than thoae

DD 1
FORM
JAN 64 1473 (BACK) 118 UNCLASSIFIED
Security Classification